[WLANware] [christof.schulze at gmx.net: babeld usage for wifi mesh for festival in Germany]
Christof Schulze
christof.schulze at gmx.net
Sat Jul 6 12:48:01 CEST 2019
Hallo zusammen,
auf mehrfachen Wunsch hin möchte ich euch nun diese Nachrichten nicht
länger vorenthalten.
Die angesprochenen drei Punkte sind bereits in der jeweiligen Software
adressiert (babeld skaliert nun nicht mehr mit quadratischem Aufwand
sondern mit linearem, mmfd lauscht nicht mehr auf dem babel socket
sondern hat seine eigene neighbour-Erkennung), wodurch die CPU-Last
deutlich sinken wird. Ich bin gespannt auf die nächsten Tests in der
Größenordnung.
viele Grüße
Christof
----- Forwarded message from Christof Schulze <christof.schulze at gmx.net> -----
Date: Thu, 4 Jul 2019 23:45:02 +0200
From: Christof Schulze <christof.schulze at gmx.net>
To: Babel users mailing list <babel-users at lists.alioth.debian.org>
Cc: User Ml Ffffm <user at wifi-frankfurt.de>, geno at fireorbit.de, ffmd-dev at netz39.de
Subject: babeld usage for wifi mesh for festival in Germany
Message-ID: <20190704214502.eowtr74zu5o5gwmh at mail.christofschulze.com>
User-Agent: NeoMutt/20170113 (1.7.2)
Hello everyone,
Babeld 1.8.4 was just used to create a mesh network for the german festival
"Breminale". To accomplish client roaming, l3roamd was used.
Multicast-Routing was enabled mesh-wide using mmfd. Before the network was
taken down and replaced with a switched network it had 5-6K Routes and 1K
client devices (with 58 Nodes with 2-4 used interfaces into every daemon on
it).
This is proof that this type of setup is able to scale into the dimensions
needed by Freifunk Communities in Germany. This is the largest real-life
babeld installation that I know of and the fact that it was possible to pull
it off is a great success. I could stop this email here, however the engineer
in myself cannot deny that:
* Route distribution was slow at that network size to the point where the
network was unusable during peak times.
* Babeld was spending a lot of CPU time. (1.9 should help)
* mmfd was listening on the babeld status socket, burning 30% CPU.
Monitoring just neighbour changes would have significantly helped. To address
this, an architecture change is being implemented in mmfd such that the
dependency towards the babel socket is removed. This will reduce the load as
well.
I am extremely curious how the stack will perform next time this is attempted
using babeld 1.9.
I thank the Freifunk Bremen team around genofire for allowing this type of
experiment at that scale and for making it all happen, providing extremely
valuably input for debugging and further developing the required software
stack [1]. Thank you very much.
[1] https://github.com/FreifunkBremen/gluon-site-ffhb/tree/breminale2019-babel
Cheers
Christof
--
() ascii ribbon campaign - against html e-mail
/\ www.asciiribbon.org - against proprietary attachments
----- End forwarded message -----
--
() ascii ribbon campaign - against html e-mail
/\ www.asciiribbon.org - against proprietary attachments
More information about the WLANware
mailing list